Multimedia Similarity Detection with Mel Spectrogram Audio Features

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing multimedia information processing systems struggle to accurately identify similarities between videos that have undergone complex editing, such as excessive cropping or audio manipulation, leading to low similarity identification accuracy and potential copyright infringement.

Innovation Solution

A multimedia information processing method involving audio feature extraction using Mel spectrograms and a VGGish network to determine audio feature vectors, followed by similarity analysis through a triplet-loss layer network, to enhance identification accuracy and detect potential copyright infringements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If video image fingerprints are used for similarity identification, then the process is simple, but the accuracy is low when videos have been excessively cropped or edited

Engineering Contradiction:
Improveidentification process complexityVSAvoidsimilarity identification accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the video into multiple audio frames and extracts audio features from each frame independently. Instead of treating the entire video as a single unit, the system divides it into manageable segments (frames) and processes each segment to extract acoustic features, which are then aggregated for similarity comparison. This segmentation approach maintains robustness against video editing while preserving accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent replaces the mechanical/video-based fingerprint system with an acoustic/audio-based system. By substituting visual analysis with auditory analysis, the system becomes insensitive to visual transformations like cropping, filtering, or visual effects, while maintaining the ability to detect similarity through audio content that remains unchanged by such edits.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If audio feature extraction is added to improve accuracy, then the identification accuracy improves, but the processing complexity increases

Engineering Contradiction:
Improvesimilarity identification accuracyVSAvoidprocessing system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements a multi-functional system that can handle both video-based and audio-based similarity identification. The same framework processes different types of features (visual and acoustic) using unified similarity comparison logic, making the system versatile and reducing overall complexity despite the added audio processing capability.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent introduces an intermediary audio feature extraction layer that bridges the original video content and the similarity comparison process. This intermediary layer converts complex video data into simplified acoustic feature representations, which then feed into the similarity identification system, reducing the complexity of direct video comparison while maintaining accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If deep metric learning with triplet loss is used, then the feature discrimination capability is enhanced, but the training data requirements and computational resources increase

Engineering Contradiction:
Improvefeature discrimination capabilityVSAvoidtraining data quantity
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent applies parameter changes by modifying the loss function parameters in the neural network training process. By adjusting the triplet loss parameters (margins, weights, and thresholds) and using cross-entropy loss with carefully selected class labels, the system achieves effective feature discrimination without requiring excessively large training datasets. The parameter optimization allows the model to learn robust acoustic features from limited training data.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4114012B1Method and apparatus for processing multimedia information, and electronic device and storage medium
Publication Date: 2025.10.22 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP4114012B1 patent drawingFigure 1
  • EP4114012B1 patent drawingFigure 2
  • EP4114012B1 patent drawingFigure 3

AI summary

This application provides a multimedia information processing method and apparatus, an electronic device, and a storage medium. The method includes: parsing multimedia information to separate an audio from the multimedia information; converting the audio to obtain a mel spectrogram corresponding to the audio; determining, according to the mel spectrogram corresponding to the audio, an audio feature vector corresponding to the audio; and determining, based on an audio feature vector corresponding to a source audio in source multimedia information and an audio feature vector corresponding to a target audio in target multimedia information, a similarity between the target multimedia information and the source multimedia information.